Search

This paper focuses on the frequently occurring issue of automotive radar sensors: ghost detection. Three data-based approaches, namely random forest, convolutional neural network (CNN), and PointNet++, are adopted to identify ghost detection. Evaluated with the same dataset, random forest and PointNet++, with more than 95% accuracy, are evidently better than CNN in not only city but also motorway scenarios. Furthermore, the influence of various features for each classifier is also analyzed.

Automotive radar is a promising technology with regard to path planning, since radar systems offer a comparatively long range and are robust against bad weather conditions. In this paper, we use Convolutional Neural Networks (CNN) to determine the current road course from radar point clouds. For this purpose, we first transform the radar point cloud into various gridmaps, which then serve as an input for the CNN. The quality of the road course estimation is evaluated using a test dataset. Exemplary test results showed an average deviation of less than 91 cm at a range of 100 m between the ground truth and the estimated road course. These excellent results prove that CNN processing of radar measurements is an attractive option for reliable and precise road course estimation.

This paper addresses one of the most serious drawbacks of automotive radar sensors: The frequent occurrence of ghost detections. As such detections may cause undesired behavior, they need to be partially or completely eliminated. We present an algorithm, which uses a machine-learning-based classifier to distinguish between real and ghost detections. In contrast to other papers, this approach addresses all causes of ghost detections and not only simple multipath scenarios. Real world experiments – also in challenging situations – show success rates of 91%.

We introduce a method to classify ghost moving detections in automotive radar sensors for advanced driver assistance systems. A fully connected network is used to distinguish between real and false moving detections in the occupancy gridmaps. By using this architecture, we combine the local Doppler information, along with the spatial context of the surrounding scenario to classify the moving detections. A proof of concept experiment shows promising results with data from a test drive in an urban scenario.

This paper introduces a generalized framework for positioning mobile nodes in a wireless network. The transformation of the range data measured between pairs of network nodes into a spatial position is a very challenging task if the range measurements are distorted. Typical distortions in wireless ranging systems are caused by an imperfect clock synchronization or by multipath reflection. The positioning method proposed in this paper overcomes problems of the traditional circular or hyperbolic methods and allows for the detection and elimination of distorted measurements. The positioning is done in a non-Markovian manner, which is an advantage over other available methods, especially the Kalman-based positioning methods, since linear assumptions in the dynamics are avoided. Another advantage of the proposed method is that the two tasks positioning and smoothing are carried out separately. Hence, smoothing of the position data, e.g. by means of a Kalman filter, is possible without the well documented problems in conventional methods induced by the usually applied simplifying assumptions of linear dynamics. The very good performance of the presented approach is demonstrated by real life results obtained in industrial environments.